Authors :
Presenting Author: Spencer Morris, MS – Cleveland Clinic
Xiaowei Xu, MD – Cleveland Clinic
Ting-Yu Su, PhD – Cleveland Clinic
Demitre Serletis, MD, PhD – Cleveland Clinic Epilepsy Center, USA
Hiroatsu Murakami, MD, PhD – Cleveland Clinic
Andreas Alexopoulos, MD, MPH – Cleveland Clinic
Stephen Jones, MD – Cleveland Clinic
Imad Najm, MD – Cleveland Clinic
Zhong Irene Wang, PhD – Cleveland Clinic
Rationale:
Focal cortical dysplasia (FCD) is a common cause of pharmacoresistant focal epilepsy in the pediatric and adult populations but is difficult to identify on conventional MRI. MRI post-processing techniques (Huppertz et al., Epilepsy Res. 2005) may aid detection but require trained human readers to interpret and has low specificity. Therefore, deep and machine learning approaches have been developed to automatically identify these lesions. Recent work shows that 3D convolutional neural networks (CNNs) perform better for FCD detection than their per-slice 2D variants (Kersting et al., Epilepsia 2025). Including 3D FLAIR also improves CNN performance (Gill et al., Neurology 2021). However, retrospective FCD cohorts often contain 2D axial and coronal FLAIR images only. We have thus developed a method for generating super-resolution (SR) 3D FLAIR images from 2D axial FLAIR, 2D coronal FLAIR, and 3D T1w inputs.
Methods: We used data from 250 subjects (134 epilepsy patients, 116 healthy controls) that had 3D FLAIR and 3D T1w images: 159 were from the University of Bonn dataset (Schuch et al., Sci. Data 2023), 16 from the IDEAS dataset (Taylor et al., Epilepsia 2024), and the rest from the Cleveland Clinic.
Lesion ROIs for all datasets were defined by expert review. 216 subjects had 3D FLAIR only; 2D axial and coronal FLAIR images were generated by downsampling in the desired through-plane with randomized slice thicknesses The remaining 34 had preexisting 2D axial and coronal FLAIR. Most epilepsy patients had FCD II; other pathologies (e.g., PVNH, MOGHE) were included for generalizability. As seen in Figure 1, we constructed a generative-adversarial network (GAN) based on ESRGAN (Wang et al., ECCV 2018) to synthesize SR 3D FLAIR. The discriminator used a pyramid-pooling architecture for multi-scale feature extraction (Wang et al., WACV 2020). Data were split into training (80%), validation (10%), and testing (10%) sets. Training was patch-based. Lesional ROIs were oversampled when available for proper lesion translation. Performance was evaluated by computing the structural similarity (SSIM) between the SR and ground truth 3D FLAIR images in the test set; this was compared to SSIM from the interpolated 2D FLAIR. SSIM was calculated based on brain voxels for all subjects and lesion voxels in lesional cases. These values were compared with
post-hoc Wilcoxon signed-rank tests.
Results: The median SSIM for the SR 3D FLAIR images was 0.96, significantly higher than that for the interpolated 2D coronal (0.81, p < 0.001) and 2D axial (0.77, p < 0.001) FLAIR. In lesional patients, the median SSIM for lesional voxels was 0.93 for the SR 3D FLAIR, again significantly greater than that for the interpolated 2D coronal (0.81, p < 0.001) and 2D axial (0.79, p < 0.001) FLAIR (Figure 2B). Subtle features of FCD II lesions were preserved (Figure 2A).